The Prophecy of the Answering Machine

A short history of the recurring computer-science dream: a machine that can answer every question.

August 11, 2026 · 5 min read

This post is a work in progress.

I love conversational analytics, it kinda feels magical when you ask a talk-to-data agent questions and it’s able to compose a SQL query or search documents to precisely answer your question.

All computer scientists have had a fantasy, and that is to build a machine that can answer every question in the universe. We’ve had multiple narrow solutions like search engines, Wikipedia, Siri etc. (todo: come up with more examples here) for this but the most significant leap came with LLMs and AI in general.

In my head, this fantasy to build this machine stems from our primal instinct to look up at the sky and ponder, and to ask questions.

The definition of AGI keeps changing, but answering questions is a subset of the current accepted definition of “being able to do all economically viable work”.

Funnily enough, if you study the history of data systems, you’ll always find traces of efforts to build this answering machine.

Year	Milestone	Research strand	What it added to the answering machine
1958	Luhn’s “Business Intelligence System”¹	Organizational information systems	Proposed an automated system that would ingest, retrieve, distribute, and furnish information on demand inside an organization.
1961	BASEBALL²	Natural-language database access	Answered restricted ordinary-English questions over stored baseball records and explicitly imagined executives, commanders, and scientists questioning computers directly.
1966	ELIZA³	Conversational interface	Demonstrated how strongly a textual dialogue could suggest intelligence, despite relying on scripted transformations rather than broad knowledge.
1970	Codd’s relational model⁴	Database architecture	Separated logical querying from physical storage, allowing future users to work with data without understanding its internal machine representation.
1973	LUNAR⁵	Natural-language database access	Let lunar geologists query Apollo sample data in ordinary English, supported by a carefully bounded vocabulary and domain model.
1974	Relational support for non-programmers⁶	Database accessibility	Made the “casual user” an explicit database-design objective and proposed formal or informal language interfaces over relational data.
1982	Chat-80⁷	Natural-language database access	Established a strikingly modern pipeline: English question → logical representation → query planning → execution → answer.
1999	TREC-8 Question Answering track⁸	Open-domain question answering	Shifted the retrieval objective from returning ranked documents to returning actual answers over a large document collection.
2003	Neural probabilistic language model⁹	Neural language modeling	Learned distributed word representations and sequence probabilities together, laying groundwork for models that acquire linguistic regularities rather than relying only on handcrafted rules.
2010–11	IBM Watson and DeepQA¹⁰	Open-domain question answering	Combined question analysis, retrieval, candidate generation, evidence scoring, and ranking at sufficient breadth and speed to defeat leading Jeopardy! champions.
2013	Power BI Q&A¹¹	Natural-language BI	Brought natural-language questions to business data while documenting its dependence on a curated data model, synonyms, data quality, and ambiguity handling.
2017	The Transformer¹²	Neural architecture	Replaced recurrence with attention in a highly parallelizable architecture, enabling the training scale behind modern large language models.
2018-2020	GPT-1¹³, GPT-2¹⁴, GPT-3¹⁵	Generative pretraining	Showed that one Transformer language model pretrained on unlabeled text could transfer to many language-understanding tasks with relatively small task-specific adaptations.
November 2022	ChatGPT¹⁶	Product convergence	Combined broad pretrained knowledge, instruction following, RLHF, dialogue-specific training, conversational state, and an accessible interface. It fulfilled the interface prophecy, while its plausible false answers revealed that the source-of-truth problem remained unsolved.

The earlier systems were narrow because their semantics were explicit: they depended on schemas, vocabularies, logical forms, join graphs, and curated data models. Large language models became broad by making much of that knowledge implicit in model weights.

Footnotes

Hans Peter Luhn, “A Business Intelligence System,” IBM Journal of Research and Development 2, no. 4 (1958): 314–319. IBM PDF. ↩︎
Bert F. Green Jr., Alice K. Wolf, Carol Chomsky, and Kenneth Laughery, “BASEBALL: An Automatic Question-Answerer,” Western Joint Computer Conference (1961): 219–224. Paper. ↩︎
Joseph Weizenbaum, “ELIZA—A Computer Program for the Study of Natural Language Communication Between Man and Machine,” Communications of the ACM 9, no. 1 (1966): 36–45. DOI. ↩︎
E. F. Codd, “A Relational Model of Data for Large Shared Data Banks,” Communications of the ACM 13, no. 6 (1970): 377–387. IBM Research. ↩︎
William A. Woods, “Progress in Natural Language Understanding: An Application to Lunar Geology,” AFIPS National Computer Conference 42 (1973): 441–450. DOI. ↩︎
E. F. Codd and C. J. Date, “Interactive Support for Non-Programmers: The Relational and Network Approaches,” SIGFIDET (1974). IBM Research. ↩︎
David H. D. Warren and Fernando C. N. Pereira, “An Efficient Easily Adaptable System for Interpreting Natural Language Queries,” American Journal of Computational Linguistics 8, nos. 3–4 (1982): 110–122. ACL Anthology. ↩︎
Ellen M. Voorhees and Dawn M. Tice, “The TREC-8 Question Answering Track Evaluation,” Eighth Text REtrieval Conference (1999): 83–105. NIST proceedings. ↩︎
Yoshua Bengio, Réjean Ducharme, Pascal Vincent, and Christian Jauvin, “A Neural Probabilistic Language Model,” Journal of Machine Learning Research 3 (2003): 1137–1155. JMLR. ↩︎
David Ferrucci et al., “Building Watson: An Overview of the DeepQA Project,” AI Magazine 31, no. 3 (2010): 59–79. IBM Research; see also IBM’s history of the 2011 Jeopardy! result. ↩︎
Microsoft Power BI Team, “Live Now! Q&A with Your Data,” December 18, 2013. Power BI Blog. ↩︎
Ashish Vaswani et al., “Attention Is All You Need,” Advances in Neural Information Processing Systems 30 (2017). NeurIPS. ↩︎
Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever, “Improving Language Understanding by Generative Pre-Training,” OpenAI, 2018. Paper. ↩︎
Alec Radford et al., “Language Models are Unsupervised Multitask Learners,” OpenAI, 2019. Paper. ↩︎
Tom B. Brown et al., “Language Models are Few-Shot Learners,” Advances in Neural Information Processing Systems 33 (2020). OpenAI. ↩︎
OpenAI, “Introducing ChatGPT,” November 30, 2022. OpenAI. ↩︎